Search Memory

Search Memory

Searches observations and dossiers across all projects. Like type:decision file:auth.ts from structured memory systems, with zero external dependencies.

---

Query Syntax

Supports filter combinations:

| Filter | Example | Searches | |--------|---------|----------| | Free text | caching | Summary and details across all observations | | type: | type:decision | By type: decision, bugfix, feature, discovery, problem | | project: | project:my-app | Within a specific project | | file: | file:auth.ts | By affected files | | date: | date:2026-02 | By date (year-month or year-month-day) | | status: | status:open | Open problems (problem without resolution) |

Example queries

/search-memory type:decision caching
/search-memory project:my-app type:problem
/search-memory file:dashboard
/search-memory type:problem status:open
/search-memory date:2026-02 type:feature
/search-memory what do we know about websockets

---

Search Algorithm

Step 1: Determine scope

• If project: is specified, search only that project
• If not specified, search all projects

Step 2: Load indexes

For each project in scope:

Read ~/.claude/memory/projects/{project}.observations.md

Parse only the Index table first (progressive disclosure: do not load Details).

Step 3: Filter by index

Apply filters to index rows:

• type: matches the Type column
• date: matches the Date column
• file: matches the Files column
• Free text matches the Summary column

Step 4: Load details (on demand)

For matched entries, load the Details section only for those matches.

This is progressive disclosure: index costs ~40 tokens per entry, details cost ~150 tokens. With 100 observations, savings: 4,000 tokens instead of 15,000.

Step 5: Also search dossiers

Additionally check {project}.md dossier files:

• "Decisions Made" section
• "Unresolved Problems" section
• "Next Steps" section

---

Output Format

## Search results: "{query}"
Found: {N} observations in {M} projects

### project-name ({n} matches)
| # | Date | Type | Summary |
|---|------|------|---------|
| 3 | 2026-01-02 | problem | Scraping too expensive |

**[3] Details:**
Before: High bills
After: Not resolved
Files: scheduler/*, worker/*

### another-project ({n} matches)
...

---

Special Queries

"All open problems"

/search-memory type:problem status:open

Finds all entries of type problem where Details does not contain "Status: Resolved".

"What do we know about X?"

/search-memory {concept}

Searches all fields: summary, details, files. Also checks dossiers.

"Decision history for project"

/search-memory project:my-app type:decision

Chronological list of all decisions with before/after context.

"What changed in file X?"

/search-memory file:dashboard_api.py

All observations affecting that file.

---

Important

• Progressive disclosure: always start with the index, load details only for matches
• Sort results by date (newest first)
• If nothing found, say so directly. Do not make things up
• Maximum 20 results. If more, show first 20 and state how many remain

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: vasilievyakov
• Source: vasilievyakov/miracle-infrastructure
• License: MIT

Install and usage instructions live in the source repository linked above.